diff --git a/agent_v0/server_v1/stream_processor.py b/agent_v0/server_v1/stream_processor.py
index 74f113de2..9b2cbcbc0 100644
--- a/agent_v0/server_v1/stream_processor.py
+++ b/agent_v0/server_v1/stream_processor.py
@@ -502,14 +502,24 @@ def _vlm_identify_element(anchor_b64: str, window_title: str = "") -> str:
                 if raw.lower().startswith(prefix.lower()):
                     raw = raw[len(prefix):]
                     break
+            # Rejeter les réponses qui sont du bavardage, pas un label
+            reject_patterns = (
+                "several", "multiple", "various", "image",
+                "I can", "there are", "there is", "elements",
+                "the following", "here are",
+            )
+            if any(p in raw.lower()[:30] for p in reject_patterns):
+                logger.debug("VLM identify : réponse bavarde rejetée (raw='%s')", raw[:60])
+                return ""
+
             # Prendre les 5 premiers mots utiles
             words = raw.split()[:5]
             label = " ".join(words).strip('",.\' ').rstrip(".")
-            if label and 2 <= len(label) <= 60:
+            if label and 2 <= len(label) <= 40:
                 logger.info("VLM identify element : '%s'", label)
                 return label
             else:
-                logger.debug("VLM identify : label trop court ou vide après nettoyage (raw='%s')", raw[:80])
+                logger.debug("VLM identify : label trop court/long après nettoyage (raw='%s')", raw[:80])
     except Exception as e:
         logger.debug("VLM identify element échoué : %s", e)